Novel Deep Learning Model with CNN and Bi-Directional
LSTM for Improved Stock Market Index Prediction
Jithin Eapen Abhishek Verma
Department of Computer Science Department of Computer Science
California State University New Jersey City University
Fullerton, California 92831 Jersey City, NJ 07305
jithin.john@csu.fullerton.edu averma@njcu.edu
Abstract— Predicting variations in stock price index has been
an important application area of machine learning research. Due
to the non-linear and complex nature of the stock market making
predictions on stock price index is a challenging and non-trivial
task. Deep learning approaches have become an important
method in modeling complex relationships in temporal data. In
this paper: (i) we propose a novel deep learning model that
combines multiple pipelines of convolutional neural network and
bi-directional long short term memory units. (ii) Proposed model
improves prediction performance by 9% upon single pipeline
deep learning model and by over a factor of six upon support
vector machine regressor model on S&P 500 grand challenge
dataset. (iii) We illustrate the improvement in prediction
accuracy while minimizing the effects of overfitting by presenting
several variations of multiple and single pipeline deep learning
models based on different CNN kernel sizes and number of bi-
directional LSTM units.
Keywords— deep learning; Bi-directional LSTM; stock market
prediction; CNN; S&P 500.
I. INTRODUCTION
Stock markets are some of the most import financial
institutions of any capitalist economy. It allows enterprises to
raise money from the investing public to fund their growth and
operations. At the same time there is a large amount of risk
involved for the investors due to its volatile nature. Hence
both private investors and financial institutions like banks,
have always attempted to devise ways to predict the changes
occurring in the stock prices and make buy and sell decisions
early.
The developments in the field of machine learning allowed
the use of computers in the task predicting financial markets,
by automating the existing statistical methods of data analysis
[1]. Machine learning is about building computer systems or
programs that can learn from data. Traditional machine
learning approaches to stock prediction have focused on
improving their performance with different techniques for
feature extraction to select the most promising features from a
dataset. The most popular is the support vector machine, a
type of kernel based supervised machine learning algorithm.
While being versatile, they are not as powerful as neural
networks, especially the modern deep learning networks. Deep
learning networks are powerful machine learning algorithms
that make use of many cascading layers in order to learn
multiple levels of representations. This corresponds to
different levels of abstraction. The levels form a hierarchy of
Doina Bein
Department of Computer Science
California State University
Fullerton, California 92831
dbein@fullerton.edu
concepts. Thereby deep learning models can achieve high
accuracy on prediction and classification tasks.
Once an accurate model is learned from training data,
users will be able to feed temporal stock price data of fixed
length to the model and be able to get the output as the
prediction for the next seven days. It can be used to predict
short term price changes, by various stakeholders like
investors, financial analysts, investment bankers, as one of the
factors affecting their investment decisions. Our primary
objective is to analyze a sequence of time-series data and give
predicted price for the next seven days. For this purpose, we
selected the Standard and Poor’s (S&P) 500 stock dataset,
publicly available on Yahoo Finance [2].
The paper is organized as follows. In Section II we survey
several representative techniques of machine learning and
neural networks that are used for stock price prediction. In
Section III we describe the dataset used for benchmarking
learned models. In Section IV we present our research
methodology and proposed novel model, which gives
improved performance in comparison with the state of the art
models. Section V presents experimental environment.
Experimental results are presented in Section VI. Concluding
remarks and future work is discussed in Section VII.
II. RELATED WORK
A. Automation of Stock Market
Stock market traders today use electronic exchange
systems to automate the trading of stocks based on theoretical
buy and sell prices, which are calculated based on the current
market price of the stock [3]. Such an automated system
provides the advantages of better response times to market
changes, more accurate trading operations, and reduced risk of
loss due to repeated or mistaken operations.
Traders have developed multiple methods for predicting
future stock prices: Fundamental stock analysis is the study of
the business value of the stock, by considering factors like
revenue, operating expenses, assets, products, management,
competitors, economic environment etc. Technical analysis is
based on the belief that markets tend to move in cycles and
accurate predictions can be made by studying the price
patterns and price trends of the past. Traders have been using
statistical tools to analyze price and volume data. Recent
advances in computing and machine learning has enabled
them to analyze larger datasets.
Figure 1. S&P 500 index closing prices from January 2, 2008 to November 27, 2018. [2]
B. Support Vector Machine Regressor
An important characteristic of stock prices is their time
series dependency, i.e., stock price at a particular time is
dependent on the price during the previous instance. In [4] the
authors proposed to use Support Vector Machine (SVM)
Regression based model to predict stock prices, as it is a
suitable learning technique for recognizing patterns in a time
series dataset. Different windowing methods were used to feed
more reliable inputs to the regression model. The accuracy of
the model depended on choosing the best parameters for the
support vector machine. This pre-processing of the input data
for rectangular window and flattened window were found to
have good prediction for stock prices 1 and 5 days ahead
respectively. This model was applied only to the time series
dataset of a single company in the stock market.
C. Artificial Neural Network
Artificial neural network (ANN) models have been
compared to other statistical models [5] and generally ANN
models give better result [6]. The authors then compare
various types and hybrids of ANN and multi-layer perceptron
models and found that hybrid adaptive neural networks need
to be remodeled when training data changes for the time-series
forecast. In addition, each layer has to be controlled by
reliability tests, as it is the input from the following layers of
regression model.
D. Deep Neural Networks
Recurrent Neural Networks (RNN) are a special form of
ANN where the weights computed during training are
influenced not only by the current training example, but by
previous decisions made on earlier training data. They have
been specifically designed to deal with sequential or temporal
data [7]. In RNNs, the hidden layer receives input not only
from its immediate predecessor (input) layer, but also from a
previous output layer. Thus, for each input term in the
sequential data, the weights calculated also depend on the
output prediction fromr the previous term. This architecture
conforms perfectly with the time-series nature of stock price
data. A hierarchy of Deep RNNs have been shown to perform
well for prediction tasks as related to time-series data [8].
One drawback to using the RNNs is that they result in long
connections between the final output layer, and hidden layers
of previous data in the sequence, thus resulting in the problem
of the vanishing gradients. To overcome this problem,
mmodifications were made to the RNN units to include gates
and memory units to remember the gradients. The most
popular such units are Long Short Term Memory (LSTM)
units [10]. LSTM units are designed to enhance the ability of
RNNs to remember previous values and prioritize learning
from newer data in a sequence upon older data. The memory
of LSTM units also solves the problem of vanishing gradients,
as the memory allows a gradient to move from one hidden
layer to the next without being reduced, letting early layers of
a neural network train almost as well as later layers [9].
LSTMs were shown to be more effective than conventional
RNNs [10].
Bi-directional Recurrent Neural Networks [11] are used to
learn from both the forward and backward time dependencies.
In a Bi-directional LSTM, each unit is split into two units
having the same input and connected to the same output. One
unit is used for the forward time sequence and the other for the
backward sequence. The common output does an arithmetic
operation like sum, average or concatenation on both the units.
Bi-directional LSTM is thus useful when learning from long
spanning time-series data.
A technique that is suitable for feature extraction on high-
dimensional data, as found in stock prices, is convolutional
neural networks [12]. A convolutional neural network (CNN)
is very similar to an ANN but makes specific assumptions
about the input data. That allows it achieve higher invariance
when encoding properties of input data into the network
 Output Layer
Figure 2. Diagram of basic structure of single pipeline CNN based Bidirectional LSTM network. We use this in our experiments as a baseline deep learning model
to compare with proposed novel multiple pipeline deep learning model.
architecture [13]. A typical CNN includes convolution layers,
rectified linear unit (ReLU) as activation function, pooling,
and fully connected layers. CNNs are the main workhorse to
solve modern computer vision and deep learning problems
ever since AlexNet achieved first place in the ILSVRC 2012,
an annual ImageNet Challenge [14]. CNNs need to be trained
on large amounts of training data to generate deep learning
models that achieve high generalization accuracy.
Generalization accuracy refers to the performance of the
models on test data. The hidden units, instead of merely
calculating a function on the input, performs a convolution
operation using a filter of weights. It is thus capable of
exploiting the time structure of the input by performing
convolutions on overlapping windows. They also use pooling
operations, which combines nearby values in input or feature
space through a max, average or histogram operator. The
purpose of pooling is to achieve invariance to small local
distortions and reduce the dimensionality of the feature space.
Most deep learning architectures are created using a single
sequence of multiple layers, each performing a specific
operation on the data. A multi-pipeline architecture, consisting
of four parallel sequences of layers, has been shown to give
lower classification errors on sentiment analysis in the text
domain as compared to a single pipeline model [15]. Each
pipeline was created using a sequence of CNN, Batch
Normalization, ReLU activation, Pooling, and LSTM layers.
The output of each pipeline is concatenated to give a single
prediction.
III. DESCRIPTION OF DATASET
The S&P 500 index (formerly Standard & Poor's 500
Index) is a market-capitalization-weighted index of the 500
largest U.S. publicly traded companies by market value. The
index is widely regarded as the best single gauge of large-cap
U.S. equities. Thereby it is an excellent measure of the
performance of overall stock market.
S&P 500 dataset is publicly available on Yahoo Finance
[2]. It consists of historical daily opening and closing stock
prices along with a few other additional market parameters.
Fig. 1 shows the daily closing price of S&P 500 index from
January 2, 2008 to November 27, 2018. It has the daily data
for five days in a week for a period of 10 years. It consists of
columns like daily Opening price, Closing price, High and
Low prices, and daily Volume of stocks traded. The S&P500
was chosen due to its stability compared to individual
company stock prices. Statistical analysis shows that the
autoregressive moving average model for S&P 500
outperforms the London Stock Exchange, suggesting better
potential for predictive models [16].
Stock index data is segmented into sequences of 50 closing
prices, with the last price of each sequence being one day after
the sequence before it. We construct multiple sequences of
length 50, and feed the deep learning model the entire batch as
an input. Each sequence is incremented by one day.
Additionally, tests were performed using sequences of length
14, 28, and 56, but the length of 50 performed the best out of
the three other lengths. Each of the data segments was
normalized by subtracting each price in the sequence by the
first price in the sequence, then divide by the first price in the
sequence. This form of normalization is called relative change.
Normalization means that the deep learning network will look
at the degree to which the prices changed over time, rather
than the raw prices themselves, allowing it to generalize to a
wider range of input values.

Figure 3. Diagram of the best performing proposed multiple pipeline model with CNNs followed by Bidirectional LSTMs.
IV. PROPOSED MULTIPLE PIPELINE CNN AND BI-DIRECTIONAL
LSTM MODEL
We constructed various deep learning architectures and
compared their performance to find the best performing
model. We also implement a Support Vector Machine
Regressor model for comparison. A SVM Regressor model
was created using Scikit-learn [20] library in python. We
perform grid search to find the best parameters for C and
gamma for a model using a grid of both radial basis function
kernel and linear kernel.
Then, we test several single pipeline deep-learning models
using using CNN and Bi-directional LSTM. We find the best
single pipeline model, configured as group of 1-dimensional
Convolutional layers, followed by two Bi-directional LSTM
layers with dropout. The CNN layers use ReLU as the
activation function and is followed by a Max-pooling layer
with pool size of 2. Fig. 2 diagram shows basic structure of
single pipeline CNN based Bi-directional LSTM network. We
use this in our experiments as a baseline deep learning model
and compare its results with support vector regressor and
proposed novel multiple pipeline deep learning model.
In our proposed model, we construct a deep learning
architecture consisting of three pipelines of separate layers,
see Fig. 3. Each pipeline consists of initial group of three 1-
dimensional CNNs with a ReLU activation followed by Max-
pooling. The CNNs are used so as to extract the higher level
features before passing data to the Bi-directional LSTM neural
layers. The group of CNNs are followed by a group of two Bi-
directional LSTM layers with dropouts. Each Bi-directional
LSTM unit learns from both the forward and backward
Output
Layer
sequence of data, and uses concatenation to merge the two
sequence outputs. Dropout layer is used to prevent overfitting
in the model. The use of LSTM also prevents the problem of
vanishing gradients.
Each pipeline takes the entire sequence of 50 prices as
input, and the CNN layers use 128 kernel windows of size 5 or
9, to return another sequence of smaller size.
The output from the CNN group is fed into the Bi-
directional LSTM layer group which:
1) learns using 200 pairs of forward and backward LSTM
units.
2) gives an output sequence of length 400
3) drops 50% (dropout) of the weights to generalize the model.
The two dimensional output of this group is then fed to a
dense layer with one unit and a linear activation, which then
gives the output for each individual pipeline.
The output of each of the three pipeline is concatenated
and given to the final layer consisting of a dense layer. The
final output is a numeric value of predicted price for seven
days into the future. As a typical regression problem, we have
used mean squared error (MSE) as the loss function to
measure the error in the predicted versus actual data. Adadelta
is the gradient descent optimization algorithm for correcting
the weights in the model.
 TABLE I TABLE III
RESULTS FROM FOUR BEST PERFORMING SUPPORT VECTOR MACHINE RESULTS FROM FOUR BEST PERFORMING PROPOSED MULTIPLE PIPELINES
REGRESSOR MODELS ON S&P 500 DATASET DEEP LEARNING MODELS WITH CNN AND BI-DIRECTIONAL LSTM ON S&P
500 DATASET
Kernel C Gamma Mean Test Mean Train
Score Score Kernel LSTM Units Mean Test Score Mean Train
Score
rbf 0.059636233 0.212095088 0.001619057 0.001367225
9 200 0.000281317 0.000204378
rbf 0.040949150 0.372759372 0.001652736 0.001349042
5 400 0.000354261 0.000284712
rbf 0.126485521 0.120679264 0.001666914 0.001444660
5 200 0.000359209 0.000234630
rbf 0.086851137 0.212095088 0.001672472 0.001433185
9 400 0.000420501 0.000345472
Note: rbf is radial basis function. Mean Test Score and Mean Train Score are
computed from Mean Squared Error. Lower values are better. Note: Mean Test Score and Mean Train Score are computed from Mean
Squared Error. Lower values are better.

V. EXPERIMENTAL ENVIRONMENT
We have developed the model using the python
programming language. We used the following libraries for
various functionalities:
• Keras [17] is a popular open-source library that enables
researchers and software engineers to rapidly define and train
many deep learning models in a short amount of time. It is
used for creating, and training neural networks. It provides a
simple interface to existing libraries like Tensorflow [18],
which allows us to use GPUs for faster training and prediction.
With Keras, we can achieve fast experimentation which is the
key component to gain insightful feedback and garner
improvements in accuracy of deep learning models for stock
prediction. Note that in Keras version 2.0, the Keras team has
removed metrics such as recall, precision, and fmeasure to
implicitly promote the use of accuracy as the main metric for
CNNs that are trained on balanced datasets [19].
• Tensorflow [18] is the backend for Keras, and facilitates
the processing of low-level operations such as tensor products
and convolutions. It is developed and maintained by Google.
validation.
• Matplotlib [21] was used for plotting the graphs for the
actual time-series as well as the predicted trends.
• Pandas [22] was used for reading value from csv files as
DataFrames.
• Numpy [23] used to perform matrix operations like flip,
reshape, and create random matrices.
All training and prediction have been performed on several
servers with multiple NVIDIA GeForce GTX 1080 Ti GPUs
for parallel processing, each GPU has 12GB of VRAM. The
servers use Ubuntu 16.10 as the operating system. Main
processor used is a collection of 4 Intel Xeon E5-2630 CPUs
@ 2.20 Ghz, with 10 cores and 256 GB RAM. Each server
uses CUDA toolkit v9.0.176 for utilizing the Nvidia GPUs.
We used two such servers with a total of 11 GPUs, allowing to
train multiple models simultaneously and consequently
drawing on the benefits from parallel architecture the
proposed multiple pipeline model.

• Scikit-learn [20] was used for performing the grid search VI. RESEARCH RESULTS AND ANALYSIS
of the model to find the best parameters using 5-fold Cross- We have used the grid search to perform five-fold cross

TABLE II
RESULTS FROM FOUR BEST PERFORMING SINGLE PIPELINE DEEP LEARNING
MODELS WITH CNN AND BI-DIRECTIONAL LSTM ON S&P 500 DATASET

Kernel LSTM Units Mean Test Score Mean Train

Score

9 200 0.000345951 0.000272510

5 200 0.000429588 0.000324599

9 400 0.000431591 0.000333719

5 400 0.000555343 0.000466796 Figure 4: Support Vector Regressor predictions on S&P 500 Dataset. X axis
shows the time segment for the index and prediction window. Y axis shows the
closing price. Continuous plot is the actual closing price, predictions are shown
Note: Mean Test Score and Mean Train Score are computed from Mean
as disjoint plots for next seven days. Data comes from years 2015 and 2016.
Squared Error. Lower values are better.

Figure 5: Single Pipeline Deep Learning Model predictions with CNN and Bi-
Directional LSTM on S&P 500 Dataset. X axis shows the time segment for the
index and prediction window. Y axis shows the closing price. Continuous plot
is the actual closing price, predictions are shown as disjoint plots for next seven
days window. Data comes from years 2015 and 2016.
validation for each of support vector regressor, single pipeline
and multiple pipeline models to find the top four performing
models and their parameter values in each of the three
aforementioned categories.
A. Support Vector Machine Regressor Models Results
Results from the four best performing Support Vector
Regressor models are shown in Table I. Five-fold cross
validation was performed along with grid search to select best
kernel, C and gamma values. Mean squared error is used as
the performance metric. Lower value for mean squared error is
considered better. From the list of rbf and linear kernels, the
grid search returned rbf kernels as the best performing models.
This was an expected result, as radial basis function kernels
generally perform better than linear kernels on non-linear data
[24]. The mean squared error observed for SVM regressor is
in the order of 10-3. From the results it is observed that there is
low variance across best four models.
B. Single Pipeline Models Results
Single pipeline deep learning model results are shown in
Table II. Various parameter values for CNN’s kernel sizes and
number of Bi-directional LSTM units were searched to obtain
Figure 6: Multiple Pipeline Deep Learning Model predictions with CNN and
Bi-Directional LSTM on S&P 500 Dataset. X axis shows the time segment for
the index and prediction window. Y axis shows the closing price. Continuous
plot is the actual closing price, predictions are shown as disjoint plots for next
seven days window. Data comes from years 2015 and 2016.
optimal parameters. Kernel sizes of 5 or 9 with Bi-directional
LSTM units as 200 or 400 resulted in four best performing
models. All mean squared error results given by the deep
learning model were in the order of 10-4.
C. Proposed Multiple Pipeline Models Results
Proposed multiple pipeline deep learning model results are
shown in Table III. CNN’s kernel sizes of 5 or 9 with Bi-
directional LSTM units as 200 or 400 resulted in four best
performing models. The kernel size of 9 for the CNN layer
was found to perform the best in combination with the Bi-
directional LSTM layers having 200 units. Mean squared error
results are in the order of 10-4.
D. Comparison of Results Across Models
Multiple and single pipeline deep learning models
outperform Support vector machine regressor models by a
factor higher than 4.7 for MSE value. Best performing
multiple pipeline model outperforms single pipeline model by
around 9%. This validates our idea that combining the three
pipelines results in improved performance, thereby the fusion
is non-redundant. Since, all of the individual pipelines in the
multiple pipelines model can be evaluated in parallel; there is
negligible difference in training time in comparison to single
pipeline model.
Best performing model for each of multiple pipelines,
single pipeline, and SVM regressor methodologies is used for
constructing prediction graphs in Figs. 4, 5, 6. X axis shows
the time segment for the index and prediction window. Y axis
shows the closing price. Continuous plot is the actual closing
price, predictions are shown as disjoint plots for next seven
days window. Predictions are based on data from past 50 days.
Data comes from years 2015 and 2016.
Predictions from the multiple and single pipelines model
Figs. 5 and 6 follow the actual prices more closely than the
support vector machine regressor model in Fig. 4. The
predicted trends in Figs. 5 and 6 give much better stock
market direction (up or down) when compared to Fig. 4. In
general deep learning models significantly outperform the
support vector machine regressor model, both in terms of
MSE and prediction trend given in Figs. 4, 5, 6.
VII. CONCLUSION AND FUTURE WORK
In this paper we proposed a new deep learning model that
combines multiple pipelines of CNN for feature extraction
from data and Bi-directional LSTM in analyzing temporal
data. We observed that CNN layers when combined with Bi-
directional LSTM show improved prediction performance
from the temporal sequence as compared to the traditional
SVM regressor. Furthermore, multiple pipelines of deep layers
that concatenate the best of the results from three individual
pipelines performs better than single pipeline model. This
demonstrates the applicability of deep learning models on
predicting both the values and trends in a time-series data. We
also presented multiple variations of our model to show how
we have increased accuracy and minimized the effects of
overfitting.
 Future research would apply deep learning toward
combining model for predicting price and trends with
sentiment analysis on news data to get even better stock price
index prediction.
REFERENCES
[1] J. Hall, G. Mani, and D. Barr, "Applying Computational Intelligence to
the Investment Process", in Proc. of 1996 CIFER: Computational
Intelligence in Financial Engineering.
[2] Yahoo Finance, S&P500 stock data [Online]. Available:
https://finance.yahoo.com/quote/%5EGSPC/history?p=%5EGSPC.
[Accessed 29 Nov. 2018].
[3] J. M. Marynowski et al., "Automated Trading System in an Electronic
Trading Exchange", U.S. Patent 7 177 833B1, Feb 13, 2007.
[4] P. Meesad and R.I Rasel, "Predicting Stock Market Price Using Support
Vector Regression," in Proc. of Int. Conf. Informatics, Electron. and
Vision, (ICIEV), 2013. doi: 10.1109/ICIEV.2013.6572570.
[5] G. Zhang, B.E. Patuwo, and M.Y. Hu, "Forecasting with artificial neural
networks:: The state of the art," Int. J. Forecasting, vol. 14, pp. 35-62,
Mar. 1998.
[6] E. Guresen, G. Kayakutlu, and T.U Daim, "Using artificial neural
network models in stock market index prediction," Expert Systems with
Applications, vol. 38, pp. 10389-10397, Aug. 2011.
[7] N. V. Bhattacharjee and E. W. Tollner, "Improving management of
windrow composting systems by modeling runoff water quality
dynamics using recurrent neural network," Ecological Modelling, vol.
339, pp. 68-76, Nov. 2016.
[8] M. Herman and B. Schrauwen, "Training and Analyzing Deep Recurrent
Neural Networks," Advances in Neural Information Processing Systems
26, (NIPS 2013).
[9] S. Hochreiter and J. Schmidhuber, "Long Short-Term Memory," Neural
Computation, vol. 9, pp. 1735-1780, 1997.
[10] W. Bao, J. Yue and Y. Rao, "A deep learning framework for financial
time," PLoS ONE, vol. 12, doi: 10.1371/journal.pone.0180944, Jul.
2017.
[11] Y.M. Schuster and K.K. Paliwal, "Bidirectional Recurrent Neural
Networks," IEEE Trans. Signal Process., vol. 45, pp. 2673-2681, Nov.
1997.
[12] CM. Langkvist, L. Karlsson, and A. Loutfi, "A review of unsupervised
feature learning and deep learning for time-series modeling," Pattern
Recognition Lett., vol. 42, pp. 11-24, Jun. 2014.
[13] W. Zhang, "Shift-invariant pattern recognition neural network and its
optical architecture," Proc. annual conf. Japan Society of Applied
Physics,1988.
[14] A.O. Russakovsky et al., "ImageNet large scale visual recognition
challenge," Int. J. Comput. Vision, vol. 115, pp. 211–252, Dec. 2015.
[15] A. Yenter and A. Verma, "Deep CNN-LSTM with Combined Kernels
from Multiple Branches for IMDb Review Sentiment Analysis," 8th
IEEE Annu. Ubiquitous Computing, Electron. & Mobile Commun.
Conf., (UEMCON), doi: 10.1109/UEMCON.2017.8249013, Oct. 2017.
[16] M. M. Rounaghi and F. N. Zadeh, "Investigation of market efficiency
and Financial Stability between S&P 500 and London Stock Exchange:
Monthly and yearly Forecasting of Time Series Stock Returns using
ARMA model," Physica A: Statistical Mechanics and Its Applicat., vol.
456, pp. 10-21, Aug. 2016..
[17] Keras: The Python Deep Learning library [Online]. Available:
https://keras.io. [Accessed 29 Nov. 2018].
[18] TensorFlow: An open source machine learning framework for everyone
[Online]. Available: https://www.tensorflow.org/. [Accessed 29 Nov.
2018].
[19] Keras Team, Keras 2.0 release notes [Online]. Available:
https://github.com/keras-team/keras/wiki/Keras-2.0-release-notes.
[Accessed 29 Nov. 2018].
[20] Scikit-Learn: Machine Learning in Python [Online]. Available:
https://scikit-learn.org/stable/. [Accessed 29 Nov. 2018].
[21] The Matplotlib development team (2018, Nov. 28). matplotlib Version
3.0.2 [Online]. Available: https://matplotlib.org/. [Accessed 29 Nov.
2018].
[22] T. Mester (2018, Jul. 10). Pandas Tutorial 1: Pandas Basics [Online].
Available: https://data36.com/pandas-tutorial-1-basics-reading-data-
files-dataframes-data-selection/. [Accessed 29 Nov. 2018].
[23] NumPy [Online], (2018). Available: http://www.numpy.org/. [Accessed
29 Nov. 2018].
[24] C.W. Hsu, CC. Chang, and C.J. Lin, "A practical Guide to Support
Vector Classification," Dept. Comp. Sci., Nat. Taiwan Univ., Taipei,
Apr. 2010